Urgent.News

What's breaking now, across thousands of outlets.

Thursday, 24 September 2026 Reading Room

AI

Supervise the agents that build your OpenSpec changes

First published on openspec-ui.dev . OpenSpec gives an agent something to build from: a proposal, a design, a list of tasks, and the spec deltas the change will leave behind. What it does not give you is a way to see the agent while it works. Several tools draw OpenSpec changes as boards and lists, and they do it well. Read on 2026-09-20 in their own words, none of the three I compared starts an…

dev.to ·

OpenSpec is a tool that provides a proposal, design, task list, and spec deltas for changes, but lacks a way to see the agent working during the process. OpenSpec Workbench addresses this by running an agent CLI on a change, providing real-time updates of what the run is doing and allowing a person to stop it if necessary. This guide offers a detailed walkthrough of how OpenSpec Workbench operates and what limitations it has.

Changes are organized as cards within the Pipeline, which runs a series of stages including propose, review, apply, verify, archive, and git. Agents are chosen for each stage, and the chain can pause for user input. Every run writes a status record, which is displayed on the card and can be printed for all runs in the repository.

Runs can occur side by side, each with its own git worktree and lease. OpenSpec Workbench allows users to stop a run where the work is sound, based on a signed request from the machine's key. The tool supports various agents such as Claude CLI, GitHub Copilot CLI, Codex CLI, Gemini CLI, DeepSeek CLI, and a local model behind an OpenAI-compatible endpoint. However, Codex and Gemini have not been tested with real binaries or over ACP.

OpenSpec Workbench enforces autonomy levels, checkpoints, review gates, and spending caps for each stage. Each run is logged and tracked in an audit log. The Workbench comes with 17 change templates for common scenarios, making it easy to start a new change with the appropriate structure. It is available for installation via VS Code or as a standalone application, and the source code and issues are hosted on GitHub.

Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at dev.to →

More in AI

The Cleaner vs Data Leaks: Sanitizing LLM Context

In Part 1 of this series , we laid out the anatomy of an agentic disaster and introduced our in-process tactical squad: The AG-Men .

  • The Cleaner prevents data leaks by sanitizing LLM context.
  • Operates on zero-trust doctrine, masking PII in memory.
  • Intercepts payloads at front door and tool boundary to enforce boundary isolation.

·dev.to

Claude Code Subagents Were 48% of My Bill. Their Output Was 0.9%

I thought my expensive Claude Code habit was long sessions. I was half right. When I finally broke down a month of usage, Claude Code subagents cost 48% of my entire bill , and the tokens they wrote…

  • Claude Code subagents comprised 48% of the budget
  • Subagents' output accounted for only 0.9% of costs
  • Fan-out workflows drove 48% of the expense

·dev.to

In 2018 I hand-wrote a C++ deep learning framework so I'd never pad a batch. In 2023 LLM serving landed on the same structure.

There's a sentence in the README of a library I wrote that I've been thinking about lately: "To summarize, we believe that Padding-free Dynamic Batching is the feature that NLPers will dive into but…

  • In 2018, master's student created C++ deep learning library InsNet
  • InsNet's design eliminated need for padded tensors by using flat buffers
  • Batching in InsNet grouped nodes by type signature, not input shape

·dev.to

More from Thursday 24 September →